ministral3¶
Ministral 3, text-only (Ministral3ForCausalLM), on vLLM.
vLLM runs Ministral3ForCausalLM through its mistral module
(MistralForCausalLM), so this is the mistral family: Llama's names,
the fused block and the classes it keys. The queries' position scaling
(llama_4_scaling_beta) is applied by vLLM's MistralAttention before
its attention layer, so attention_queries are the scaled ones, as on
transformers; vLLM reads it from a llama_4_scaling entry that only its
Mistral-format config loader writes, and the factor is exactly 1 below
original_max_position_embeddings (16384 on the released checkpoints),
so the two engines agree on any prompt shorter than that.